2023 年的期刊文章 · 第 9 页

共 1511 篇,按发表时间排,新的在前,每页 40 篇。这一页是 2023-09-01 到 2023-08-20。

按期刊、按主题看 →

第 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 页

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-09-01

David Rossell's contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

David Rossell

摘要

Congratulations on a thought-provoking piece. Building Bayesian inference from a (likelihood, predictive) pair, rather than a (likelihood, prior), enriches the paradigm and provides new ways to think about, formulate, and solve problems. A few respectful remarks. First, although the authors never claim this, it is worth emphasising that the framework is not prior-free. There is a posterior and a likelihood, hence the prior is proportional to their ratio. The key is that said prior is data-dependent, providing an interesting avenue to develop objective Bayes methods, at the cost of losing the coherence property in belief updating. Inspecting the prior can be informative. Figure 1 shows a Bernoulli example where truly θ=0.5 but the implied prior places little mass around that value, and a Gaussian example where the prior is centred around the sample mean.1 This apparently erratic prior behaviour might be problematic for model choice via Bayes factors, e.g. returning a very small integrated likelihood in the Bernoulli example. Likelihood, posterior, and prior densities for Bernoulli(0.5) and Normal(5,1) examples with n=100⁠. Second, while sometimes it is easier to elicit a predictive than a prior, in my experience the reverse is often true. For example, in regression, a prior on parameters defines a prior on the R2 coefficient, an easy-to-interpret quantity, whereas eliciting predictives may be less intuitive for nonstatisticians. Further, note that computational considerations elegantly discussed by the authors severely restrict the range of predictives one may consider in practice, limiting the flexibility of the framework. Third, I am afraid I disagree on the frameworks’ computational convenience. Doing a single optimisation may be faster than sampling, but the framework requires solving many optimisations. This is not cheaper than posterior sampling in a standard (likelihood, prior) construction, also the latter offers fast nonsampling based tools, e.g. Laplace approximations and extensions. It would be interesting to consider analogues for the predictive framework. Finally, a remark on assuming that at n=∞ there is no uncertainty left. In some settings this is not true, e.g. in high-dimensional regression with p≫n (one adds higher-order polynomial terms as n grows, say) and a normal prior on the parameters there remains posterior uncertainty even as n→∞⁠. The proposed framework does not account for such uncertainty, unless suitable adjustments are made. *(abstract via OpenAlex)*

SiM Survival Analysis · 发表于 2023-08-31

Maximum approximate likelihood estimation in accelerated failure time model for interval‐censored data

Zhong Guan

摘要

The approximate Bernstein polynomial model, a mixture of beta distributions, is applied to obtain maximum likelihood estimates of the regression coefficients, the baseline density and the survival functions in an accelerated failure time model based on interval censored data including current status data. The estimators of the regression coefficients and the underlying baseline density function are shown to be consistent with almost parametric rates of convergence under some conditions for uncensored and/or interval censored data. Simulation shows that the proposed method is better than its competitors. The proposed method is illustrated by fitting the Breast Cosmetic and the HIV infection time data using the accelerated failure time model.

JRSSB Bayesian Methods Nonparametric & Semiparametric 讨论 / 评论 · 发表于 2023-08-31

Kolyan Ray and Botond Szabo's contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Kolyan Ray, Botond Szabó

摘要

We congratulate the authors for their thought-provoking article, which includes proposing a copula-based update for the predictive density and establishing its frequentist consistency under relatively mild assumptions. In this contribution, we further explore the frequentist properties of (Bayesian) predictive densities and illustrate through a simple example that, similarly to the posterior distribution, the predictive distribution can also be inconsistent. One therefore requires caution when using martingale posteriors, at least for the frequentist. Consider a modified version of Example 1 in the present paper, taken from Christensen (2009). Let Y1,…,Yn∼iidfθ⁠, where i.e. for rational parameter values θ, we replace the N(θ,1) Gaussian density by a Cauchy with location parameter θ. As in Example 1, we endow θ with a standard Gaussian prior, i.e. π(θ)=N(θ∣0,1)⁠. Since the likelihoods in our example and Example 1 are equal almost everywhere under the prior, one can show the corresponding posteriors are also identical, i.e. the posterior is N(θ∣θ¯n,σ¯n2) with θ¯=∑i=1nyi/(n+1) and σ¯n2=1/(n+1)⁠, see Christensen (2009). Similarly, the posterior predictive is p(y∣y1:n)=N(y∣θ¯n,σ¯n2+1)⁠. By Example 2 in Hahn et al. (2018), the predictive updates can thus be characterised via a Gaussian copula with correlation parameter ρn=(1+n)−1⁠. However, for any fixed θ∈Q⁠, which forms a dense set in the parameter space R⁠, the data is Cauchy. The posterior is thus inconsistent for any rational ‘true’ parameter θ∈Q⁠, and the predictive density differs substantially from the true Cauchy density, even as n→∞⁠. Following the notation of the paper, this procedure cannot consistently recover θ∞=θ(Y1:∞)=fθ⁠, even though this parameter is fully defined by the infinite observations Y1:∞⁠. Of course, this example does not contradict Theorem 7, as the assumption ‖f0/p0‖∞≤B does not hold when f0 is Cauchy and p0 is Gaussian. In this simple example, the discontinuous likelihood function causes the posterior inconsistency, which in turn implies inconsistency of the predictive distribution. The issue is that the above approach only works for parameter values in a set of prior probability one, namely R∖Q⁠. However, prior null sets can be very large if not judged from the prior perspective. This problem becomes more pronounced in nonparametric models, where the set of parameters over which the posterior is consistent can be topologically negligible compared to those where it is inconsistent, see the classical results (Diaconis & Freedman, 1986; Freedman & Diaconis, 1983). This in turn results in predictive distributions not resembling the true data generating distribution. One must therefore be careful with the choice of predictive distribution, justifying the approach as is done in Theorem 7. We second the authors’ view that it is of interest to derive new conditions, analogous to the Kullback–Leibler property in the classical Bayesian setting, which yield consistency and ideally minimax concentration rates. Cofounded by the European Union (ERC, BigBayesUQ, project number: 101041064). Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. *(abstract via OpenAlex)*

JRSSB Bayesian Methods Nonparametric & Semiparametric 讨论 / 评论 · 发表于 2023-08-31

Isadora Antoniano Villalobos's contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Isadora Antoniano-Villalobos

摘要

I thank the authors for their refreshing ideas, providing many insights and intuitions on the process of statistical learning. For years, Bayesians have considered two levels of uncertainty, the likelihood and the prior. By shifting the focus to the predictive distribution as a single expression of uncertainty about a future observation (given data), the sequential nature of statistical learning is made explicit, together with some of its limitations. If each observation carries a piece of information, it follows that any question requiring more than (at most) a countable number of pieces cannot be answered by this process and must, rather, be assumed. This encapsulates in a single idea the need for smooth link functions and kernels, sparsity priors, and other tools used in different contexts to eliminate by design what cannot be learned from data. At the same time, assumptions such as exchangeability may lose relevance. For example, if one wished to learn the upper bound of a certain quantity, in a standard setup, repeated measurements would be assumed exchangeable. However, if measurements 1, 1.1⁠, 2.3 were observed, how much we learn from the third observation (the change in predictive uncertainty) would arguably depend on the order in which the three values were observed. The symmetry relation between measurements is not preserved by the learning process. This indicates that model definition in terms of the predictive sequence, by asking a different question, may allow inference in the presence of more complex forms of data dependence. The definition of a coherent predictive model appears challenging and the empirical distribution is the simplest possible choice. However, if only x1 was available, setting X2=x1 almost surely is hardly a reasonable way to model uncertainty. Mixture models with data-dependent weights and/or kernel parameters might be an alternative worth exploring. Rather than simply achieving continuity or greater flexibility, the choice of the kernel could be driven by the need to better extract (or filter) the information that each new observation contains about possible future samples. This idea is wonderfully illustrated by the bivariate copula construction proposed in Section 4. The copula can be interpreted as a measure of the ‘speed’ at which the predictive is updated (moves) towards its limit (ideally the distribution of interest) as new observations arrive. Studying the relationship between mixtures and copulas could provide a tool to elicit new (perhaps approximate but computationally efficient) predictive models. *(abstract via OpenAlex)*

SiM High-Dimensional Networks & Graphs Bayesian Methods · 发表于 2023-08-31

Inconsistency identification in network meta‐analysis via stochastic search variable selection

Georgios Seitidis, Stavros Nikolakopoulos, Ioannis Ntzoufras, Dimitris Mavridis

摘要

The reliability of the results of network meta‐analysis (NMA) lies in the plausibility of the key assumption of transitivity. This assumption implies that the effect modifiers' distribution is similar across treatment comparisons. Transitivity is statistically manifested through the consistency assumption which suggests that direct and indirect evidence are in agreement. Several methods have been suggested to evaluate consistency. A popular approach suggests adding inconsistency factors to the NMA model. We follow a different direction by describing each inconsistency factor with a candidate covariate whose choice relies on variable selection techniques. Our proposed method, stochastic search inconsistency factor selection (SSIFS), evaluates the consistency assumption both locally and globally, by applying the stochastic search variable selection method to determine whether the inconsistency factors should be included in the model. The posterior inclusion probability of each inconsistency factor quantifies how likely is a specific comparison to be inconsistent. We use posterior model odds or the median probability model to decide on the importance of inconsistency factors. Differences between direct and indirect evidence can be incorporated into the inconsistency detection process. A key point of our proposed approach is the construction of a reasonable “informative” prior concerning network consistency. The prior is based on the elicitation of information derived historical data from 201 published network meta‐analyses. The performance of our proposed method is evaluated in two published network meta‐analyses. The proposed methodology is publicly available in an R package calledssifs, published on CRAN and developed and maintained by the authors of this work.

SiM Bayesian Methods · 发表于 2023-08-31

Bayesian record linkage with variables in one file

Gauri Kamat, Mingyang Shan, Roee Gutman

摘要

In many healthcare and social science applications, information about units is dispersed across multiple data files. Linking records across files is necessary to estimate the associations of interest. Common record linkage algorithms only rely on similarities between linking variables that appear in all the files. Moreover, analysis of linked files often ignores errors that may arise from incorrect or missed links. Bayesian record linking methods allow for natural propagation of linkage error, by jointly sampling the linkage structure and the model parameters. We extend an existing Bayesian record linkage method to integrate associations between variables exclusive to each file being linked. We show analytically, and using simulations, that the proposed method can improve the linking process, and can result in accurate inferences. We apply the method to link Meals on Wheels recipients to Medicare enrollment records.

SiM Spatial & Time Series Bayesian Methods High-Dimensional · 发表于 2023-08-31

A Bayesian multistage spatio‐temporally dependent model for spatial clustering and variable selection

Shaopei Ma, Keming Yu, Man‐lai Tang, Jianxin Pan, Wolfgang Karl Härdle, Maozai Tian

摘要

In spatio‐temporal epidemiological analysis, it is of critical importance to identify the significant covariates and estimate the associated time‐varying effects on the health outcome. Due to the heterogeneity of spatio‐temporal data, the subsets of important covariates may vary across space and the temporal trends of covariate effects could be locally different. However, many spatial models neglected the potential local variation patterns, leading to inappropriate inference. Thus, this article proposes a flexible Bayesian hierarchical model to simultaneously identify spatial clusters of regression coefficients with common temporal trends, select significant covariates for each spatial group by introducing binary entry parameters and estimate spatio‐temporally varying disease risks. A multistage strategy is employed to reduce the confounding bias caused by spatially structured random components. A simulation study demonstrates the outperformance of the proposed method, compared with several alternatives based on different assessment criteria. The methodology is motivated by two important case studies. The first concerns the low birth weight incidence data in 159 counties of Georgia, USA, for the years 2007 to 2018 and investigates the time‐varying effects of potential contributing covariates in different cluster regions. The second concerns the circulatory disease risks across 323 local authorities in England over 10 years and explores the underlying spatial clusters and associated important risk factors.

SMMR Missing Data & Measurement Error Diagnostics & Prediction · 发表于 2023-08-30

Weighted mean difference statistics for paired data in the presence of missing values

Yuntong Li, Brent J Shelton, William St Clair, Heidi L Weiss, John L Villano, Arnold J Stromberg, Chi Wang, Li Chen

摘要

Missing data is a common issue in many biomedical studies. Under a paired design, some subjects may have missing values in either one or both of the conditions due to loss of follow-up, insufficient biological samples, etc. Such partially paired data complicate statistical comparison of the distribution of the variable of interest between the two conditions. In this article, we propose a general class of test statistics based on the difference in weighted sample means without imposing any distributional or model assumption. An optimal weight is derived from this class of tests. Simulation studies show that our proposed test with the optimal weight performs well and outperforms existing methods in practical situations. Two cancer biomarker studies are provided for illustration.

SMMR Clinical Trials · 发表于 2023-08-30

Sample sizes required to estimate the protective efficacy of a vaccine when there is an unequal allocation of individuals across the vaccine and placebo groups

Meghna Bose, Atanu Biswas

摘要

The effectiveness of a vaccine is measured by means of protective vaccine efficacy, defined by [Formula: see text], where [Formula: see text] and [Formula: see text] are, respectively, the disease attack rates in the vaccinated and the unvaccinated population. For each of the cohoret and case-control designs, methods have been presented in the literature for calculating the required sample size when the desired width of the confidence interval and the probability of coverage are pre-specified, where an equal number of individuals were assumed to be allocated to the vaccine and placebo group. In this article, we present a method for calculating the required sample size with a specified degree of precision when there is an unequal allocation of individuals across the two groups. The sample size required to achieve a desired power for the relevant level [Formula: see text] test has also been explored, keeping the unequal allocation proportion in mind. The fraction of individuals allocated to the placebo group ([Formula: see text]) can be so chosen that the total sample size or the expected number of people developing the disease or some other criteria of interest is minimized.

SMMR Missing Data & Measurement Error Survival Analysis · 发表于 2023-08-30

Multiple imputation for non-monotone missing not at random data using the no self-censoring model

Boyu Ren, Stuart R Lipsitz, Roger D Weiss, Garrett M Fitzmaurice

摘要

Although approaches for handling missing data from longitudinal studies are well-developed when the patterns of missingness are monotone, fewer methods are available for non-monotone missingness. Moreover, the conventional missing at random assumption—a natural benchmark for monotone missingness—does not model realistic beliefs about the non-monotone missingness processes (Robins and Gill, 1997). This has provided the impetus for alternative non-monotone missing not at random mechanisms. The “no self-censoring” model is such a mechanism and assumes the probability an outcome variable is missing is independent of its value when conditioning on all other possibly missing outcome variables and their missingness indicators. As an alternative to “weighting” methods that become computationally demanding with increasing number of outcome variables, we propose a multiple imputation approach under no self-censoring. We focus on the case of binary outcomes and present results of simulation and asymptotic studies to investigate the performance of the proposed imputation approach. We describe a related approach to sensitivity analysis to departure from no self-censoring. We discuss the relationship between missing at random and no self-censoring and prove that one is not a special case of the other. Finally, we discuss extensions to non-binary data settings. The proposed methods are illustrated with application to a substance use disorder clinical trial.

SiM High-Dimensional Bayesian Methods Spatial & Time Series · 发表于 2023-08-30

Detection of sparse differential dependent functional brain connectivity

Nairita Ghosal, Sanjb Basu, Dulal Bhaumik

摘要

Functional brain connectivity analysis is an increasingly important technique in neuroscience, psychiatry, and autism research. Functional connectivity can be measured by considering co‐activation of brain regions in resting‐state functional magnetic resonance imaging (rs‐fMRI). We propose a novel Bayesian model to detect differential connections in cross‐correlated functional connectivity between region of interest (ROI) pairs. The proposedsparse clustered neighborhoodmodel induces a lower‐dimensional sparsity and clustering based on a nonparametric Bayesian approach to model sparse differentially connected ROI pairs. Second, it induces a structured dependence model for modeling potential dependence among ROI pairs. We demonstrate Bayesian inference and performance of the proposed model in simulation studies and compare with a standard model. We utilize the proposed model to contrast functional connectivities between participants with autism spectrum disorder and neurotypical participants using cross‐correlated rs‐fMRI data from four sites of the Autism Brain Image Data Exchange.

SiM Spatial & Time Series · 发表于 2023-08-30

Comparison of combination methods to create calibrated ensemble forecasts for seasonal influenza in the U.S

Nutcha Wattanachit, Evan L. Ray, Thomas C. McAndrew, Nicholas G. Reich

摘要

The characteristics of influenza seasons vary substantially from year to year, posing challenges for public health preparation and response. Influenza forecasting is used to inform seasonal outbreak response, which can in turn potentially reduce the impact of an epidemic. The United States Centers for Disease Control and Prevention, in collaboration with external researchers, has run an annual prospective influenza forecasting exercise, known as the FluSight challenge. Uniting theoretical results from the forecasting literature with domain‐specific forecasts from influenza outbreaks, we applied parametric forecast combination methods that simultaneously optimize model weights and calibrate the ensemble via a beta transformation and made adjustments to the methods to reduce their complexity. We used the beta‐transformed linear pool, the finite beta mixture model, and their equal weight adaptations to produce ensemble forecasts retrospectively for the 2016/2017, 2017/2018, and 2018/2019 influenza seasons in the U.S. We compared their performance to methods that were used in the FluSight challenge to produce the FluSight Network ensemble, namely the equally weighted linear pool and the linear pool. Ensemble forecasts produced from methods with a beta transformation were shown to outperform those from the equally weighted linear pool and the linear pool for all week‐ahead targets across in the test seasons based on average log scores. We observed improvements in overall accuracy despite the beta‐transformed linear pool or beta mixture methods' modest under‐prediction across all targets and seasons. Combination techniques that explicitly adjust for known calibration issues in linear pooling should be considered to improve probabilistic scores in outbreak settings.

SiM Spatial & Time Series Nonparametric & Semiparametric High-Dimensional · 发表于 2023-08-30

Change point detection for high dimensional data via kernel measure with application to human aging brain data

Jinjuan Wang, Na Li, Zhen Meng, Qizhai Li

摘要

Identifying the existence and locations of change points has been a broadly encountered task in many statistical application areas. The existing change point detection methods may produce unsatisfactory results for high‐dimensional data since certain distributional assumptions are made on data, which are hard to verify in practice. Moreover, some parameters (such as the number of change points) need to be estimated beforehand for some methods, making their powers sensitive to these values. Here, we propose a kernel‐based ‐statistic to identify change points (KUCP) for high dimensional data, which is free of distributional assumptions and sup‐parameter estimations. Specifically, we employ a kernel function to describe similarities among the subjects and construct a ‐statistic to test the existence of change point for a given location. The asymptotic properties of the ‐statistic are deduced. We also develop a procedure to locate the change points sequentially via a dichotomy algorithm. Extensive simulations demonstrate that KUCP has higher sensitivity in identifying existence of change points and higher accuracy in locating these change points than its counterparts. We further illustrate its practical utility by analyzing a gene expression data of human brain to detect the time point when gene expression profiles begin to change, which has been reported to be closely related with aging brain.

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-08-29

Ramses Mena Chavez's contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Ramsés H Mena

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-08-29

Priyantha Wijayatunga’s contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes, and Walker

Priyantha Wijayatunga

摘要

Statistical uncertainty in (an estimate of) a parameter of a probability distribution is due to missing (unseen) observations (when it is estimated), as authors have noted. We can think that an estimate of the parameter has the maximal uncertainty when no observation is used for it, and no uncertainty when all possible observations are used for it. For example, in Bayesian sense, for a Bernoulli parameter, Beta distribution with parameter values α=1 and β=1 represents the full uncertainty. If observation counts are infinite, i.e. α and β are infinite, then there is no uncertainty. The uncertainty of the parameter estimate may be vanished when imputed or really observed data counts used for it are infinite, but the two estimates may converge to different values where the latter is the true value. But unfortunately we often do not have the chance to get it. So, it is not possible to eliminate the uncertainty correctly, i.e. while obtaining the true limiting value for the estimate, by imputing some observations given that some other observations are unknown. This is because if and only if Yn+1:∞′ is conditionally independent of Y1:n′ given the value θ1:n, which is the authors’ equation (1.1). That is, imputed observations yn+1:∞′ show only the variation that is dictated by the observations y1:n through θ1:n⁠. However, real observations yn+1:∞ from the random variable Y may show somewhat different variation unless the observations y1:n (through θ1:n⁠) determine the true probability distribution of Y. Generally we may get two different values for θ(Y1:∞) and θ(Y1:N∪YN+1:∞′) for any finite N, even though both of them are without any uncertainty. The problem is that if there are any mathematical operations or formulae for making these two values the same. Eliminating the uncertainty is not sufficient when we do not obtain the true value of the parameter. Note that, p(θ1:n|y1:n) is true posterior distribution of the parameter in the event of observed data y1:n⁠. One should use p(θ0) as the prior distribution in the Bayesian updating p(θ1:n|y1:n)∝p(y1:n|θ0)p(θ0) to make the operation clear. The authors replied later, in writing, as follows: *(abstract via OpenAlex)*

SiM Survival Analysis Nonparametric & Semiparametric · 发表于 2023-08-29

Analysis of secondary failure time responses in studies with response‐dependent sampling schemes

Yujie Zhong, Richard J. Cook, Aiai Yu

摘要

Response‐dependent sampling is routinely used as an enrichment strategy in the design of family studies investigating the heritable nature of disease. In addition to the response of primary interest, investigators often wish to investigate the association between biomarkers and secondary responses related to possible comorbidities. Statistical analysis regarding genetic biomarkers and their association with the secondary outcome must address the biased sampling scheme involving the primary response. In this article, we develop composite likelihoods and two‐stage estimation procedures for such secondary analyses in which the within‐family dependence structure for the primary and secondary outcomes is modeled via a Gaussian copula. The dependence among responses within family members is modeled based on kinship coefficients. Auxiliary data from independent individuals are exploited by augmenting the composite likelihoods to increase precision of marginal parameter estimates and enhance the efficiency of estimators of the dependence parameters. Simulation studies are carried out to evaluate the finite sample performance of the proposed method, and an application to a motivating family study in psoriatic arthritis is given for illustration.

SiM Bayesian Methods Genetics & Omics · 发表于 2023-08-29

A Bayesian method for estimating gene‐level polygenicity under the framework of transcriptome‐wide association study

Arunabha Majumdar, Bogdan Pasaniuc

摘要

Polygenicity refers to the phenomenon that multiple genetic variants have a nonzero effect on a complex trait. It is defined as the proportion of genetic variants with a nonzero effect on the trait. Evaluation of polygenicity can provide valuable insights into the genetic architecture of the trait. Several recent works have attempted to estimate polygenicity at the single nucleotide polymorphism level. However, evaluating polygenicity at the gene level can be biologically more meaningful. We propose the notion of gene‐level polygenicity, defined as the proportion of genes having a nonzero effect on the trait under the framework of a transcriptome‐wide association study. We introduce a Bayesian approachgenepolyto estimate this quantity for a trait. The method is based on spike and slab prior and simultaneously estimates the subset of non‐null genes. Our simulation study shows thatgenepolyefficiently estimates gene‐level polygenicity. The method produces a downward bias for small choices of trait heritability due to a non‐null gene, which diminishes rapidly with an increase in the genome‐wide association study (GWAS) sample size. While identifying the subset of non‐null genes,genepolyoffers a high level of specificity and an overall good level of sensitivity—the sensitivity increases as the sample size of the reference panel expression and GWAS data increase. We applied the method to seven phenotypes in the UK Biobank, integrating expression data. We find height to be the most polygenic and asthma to be the least polygenic.

Biometrics Causal Inference Clinical Trials · 发表于 2023-08-28

Optimizing Treatment Allocation in Randomized Clinical Trials by Leveraging Baseline Covariates

Wei Zhang, Zhiwei Zhang, Aiyi Liu

摘要

We consider the problem of optimizing treatment allocation for statistical efficiency in randomized clinical trials. Optimal allocation has been studied previously for simple treatment effect estimators such as the sample mean difference, which are not fully efficient in the presence of baseline covariates. More efficient estimators can be obtained by incorporating covariate information, and modern machine learning methods make it increasingly feasible to approach full efficiency. Accordingly, we derive the optimal allocation ratio by maximizing the design efficiency of a randomized trial, assuming that an efficient estimator will be used for analysis. We then expand the scope of optimization by considering covariate-dependent randomization (CDR), which has some flavor of an observational study but provides the same level of scientific rigor as a standard randomized trial. We describe treatment effect estimators that are consistent, asymptotically normal, and (nearly) efficient under CDR, and derive the optimal propensity score by maximizing the design efficiency of a CDR trial (under the assumption that an efficient estimator will be used for analysis). Our optimality results translate into optimal designs that improve upon standard practice. Real-world examples and simulation results demonstrate that the proposed designs can produce substantial efficiency improvements in realistic settings.

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-08-28

Jiaqi Gu and Guosheng Yin’s contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Jiaqi Gu, Guosheng Yin

摘要

We congratulate the authors for their thought-provoking paper, especially in the three aspects as follows. Instead of imposing a prior π(θ) on the parameter θ of a likelihood function fθ(y)⁠, where both π and fθ are elicited subjectively, the authors proposed the framework of martingale posterior to directly model the predictive p(yn+1:∞∣y1:n) for which only the prior predictive p(y) needs to be specified. As a result, Bayesian inference can be conducted on any parameter (or statistical functional) of the true sampling distribution. As the hyperparameter ρ of the copula is chosen in such a data-driven way that p(y1:n) fits the observed data y1:n well, the impact of the initial guess for the prior predictive, p0(y)⁠, on p(yn+1:∞∣y1:n) can be adjusted properly. Unlike existing MCMC methods which generate autocorrelated posterior samples, the proposed predictive resampling algorithm is GPU-friendly and parallelisable. Thus, independent posterior samples can be obtained for inference, leading to improvement of both computational and statistical efficiency. The proposed martingale posterior framework, we believe, has the further potentials in several directions. Because variable selection can generally improve statistical efficiency of regression analysis, it is of interest to investigate how Bayesian variable selection can be conducted with control of the Bayesian false discovery rate under the framework of martingale posterior. Although a default choice is suggested for the initial guess p0(y)⁠, it is possible to leverage some information from the observed data y1:n to initialise p0(y)⁠, similar to empirical Bayes methods. This raises a question on how the martingale posterior connects to nonparametric empirical Bayes, which also estimates the prior from the data. *(abstract via OpenAlex)*

Biometrics Causal Inference Functional Data Networks & Graphs · 发表于 2023-08-28

Functional Bayesian Networks for Discovering Causality from Multivariate Functional Data

Fangting Zhou, Kejun He, Kunbo Wang, Yanxun Xu, Yang Ni

摘要

Multivariate functional data arise in a wide range of applications. One fundamental task is to understand the causal relationships among these functional objects of interest. In this paper, we develop a novel Bayesian network (BN) model for multivariate functional data where conditional independencies and causal structure are encoded by a directed acyclic graph. Specifically, we allow the functional objects to deviate from Gaussian processes, which is the key to unique causal structure identification even when the functions are measured with noises. A fully Bayesian framework is designed to infer the functional BN model with natural uncertainty quantification through posterior summaries. Simulation studies and real data examples demonstrate the practical utility of the proposed model.

JRSSB Bayesian Methods Nonparametric & Semiparametric 讨论 / 评论 · 发表于 2023-08-28

Blake Moya's contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Blake Moya

摘要

I would like to congratulate the authors on an expository and intuitive representation of statistical uncertainty. To assist in the further investigation of predictive resampling techniques, I have developed a software package for R (R Core Team, 2022) which implements some of the algorithms presented here as well as from subsequent work (Moya & Walker, 2023). The CopRe package (Moya, 2022), named for the copula resampling technique described in Section 4, can be installed from the Comprehensive R Archive Network with the command: install.packages(‘copre’). The copula resampling algorithm is massively parallelisable, and the simplicity of each recursive update makes implementation in very low-level programming languages quite easy. I have developed CopRe’s core code in C++ (ISO, 2012), parallelised with OpenMP (Chandra et al., 2001). I have also written core code in CUDA (NVIDIA et al., 2020) for running the algorithm on a GPU that is available upon request. A comparison of the running time for the marginal Dirichlet Process Mixture Model (DPMM) sampler of Escobar and West (1994) and Copula Resampling run in serial or parallelised over a CPU or a GPU is shown in Figure 1. The acceleration of nonparametric Bayesian inference presented by the authors is significant. Speedup over the Dirichlet Process Mixture MCMC sampler of Escobar and West (1994) in concert with the sequence resampling approach of Moya and Walker (2023) for the MCMC sampler without resampling extension and three launch configurations for CopRe. The sample size was n=100⁠, k=1000 samples were drawn from each algorithm, and for CopRe N=100 recursive predictive draws were made for each sample. Computations were made with core C++/CUDA code on an Intel Core i5 8600 K clocked to 4.8 GHz and an NVIDIA GTX 1070 Ti. By imposing prior information on the mechanics of the data-generating process rather than on its parameters, predictive resampling of martingale posteriors overcomes many of the difficulties involved with the creation of nonparametric priors, and the implementations of Markov chain Monte Carlo samplers for their corresponding posteriors. The current development version of CopRe contains a new Gibbs-type sequence resampling function, SeqRe, which exploits the known predictive update rule of many Gibbs-type priors to sample full random distributions from mixture models like the DPMM without a known representation of the prior or posterior on the random measure. The development version of the package containing new experimental features can also be installed via the command: I encourage experimenters to take advantage of this software and hope that it will accelerate further investigation of martingale posteriors. *(abstract via OpenAlex)*

SiM Causal Inference · 发表于 2023-08-27

Subgroup analysis using Bernoulli‐gated hierarchical mixtures of experts models

Wei Li, Shanshan Luo, Yangbo He, Zhi Geng

摘要

Summary: When it is suspected that the treatment effect may only be strong for certain subpopulations, identifying the baseline covariate profiles of subgroups who benefit from such a treatment is of key importance. In this paper, we propose an approach for subgroup analysis by firstly introducing Bernoulli‐gated hierarchical mixtures of experts (BHME), a binary‐tree structured model to explore heterogeneity of the underlying distribution. We show identifiability of the BHME model and develop an EM‐based maximum likelihood method for optimization. The algorithm automatically determines a partition structure with optimal prediction but possibly suboptimal in identifying treatment effect heterogeneity. We then suggest a testing‐based postscreening step to further capture effect heterogeneity. Simulation results show that our approach outperforms competing methods on discovery of differential treatment effects and other related metrics. We finally apply the proposed approach to a real dataset from the Tennessee's Student/Teacher Achievement Ratio project.

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-08-27

Proposer of the vote of thanks to Fong, Holmes and Walker and contribution to the Discussion of ‘Martingale Posterior Distributions’

Philip Dawid

摘要

Philip Dawid; Proposal of Vote of Thanks for “Martingale Posterior Distributions” by Edwin Fong, Chris Holmes and Stephen G. Walker, Journal of the Royal Statis *(abstract via OpenAlex)*

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-08-27

Pietro Rigo's contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Pietro Rigo

摘要

This paper is really interesting and offers a number of intriguing hints. Here, I just make a few isolated remarks without any claim of being exhaustive. I would give more emphasis to the Ionescu-Tulcea theorem (ITT). This theorem is quoted only in passing, but it is the cornerstone of this paper. By ITT, the distribution of Y1:∞=(Y1,Y2,…) is completely determined by the assignment of {Pn:n≥0}⁠, where P0(⋅)=P(Y1∈⋅) and Pn(⋅)=P(Yn+1∈⋅∣Y1:n)⁠. Exploiting ITT has at least two advantages: (i) The first part of the paper can be made shorter and clearer; (ii) Relying on ITT makes transparent that, in general, to introduce the problem investigated in this paper, there is no need of any distributional assumption on Y1:∞⁠. In particular, Y1:∞ needs not be exchangeable. Exchangeability should be assumed if (and only if) the inferrer feels that it is reasonable for the specific data at hand, which is true in some problems but false in others. As regards this paper, the only advantage of exchangeability is that the distribution of Y1:∞ can be assigned in two ways: by the usual likelihood/prior scheme (thanks to de Finetti’s theorem) or via ITT selecting {Pn:n≥0}⁠. These two routes are equivalent and both determine the distribution of Y1:∞⁠. Hence, it is obvious that predictive resampling is identical to posterior sampling for exchangeable data. I realise that the existence of these two routes is expository useful. But, I do not see any other general reason for assuming exchangeability from the outset. See e.g. Berti et al. (2021, 2023). Suppose the distribution of Y1:∞ is assigned via ITT, but, for some reason, Y1:∞ is requested to satisfy some distributional assumption. For instance, Y1:∞ is asked to be exchangeable, or c.i.d., or stationary, and so on. This puts some constraints on the predictive distributions Pn⁠. So, the problem arises: Is it possible to characterise a distributional assumption on Y1:∞ in terms of the Pn ? This issue has been addressed in some cases (exchangeability and c.i.d.) but not in others (stationarity, partial exchangeability). See Berti et al. (2021, 2023) and references therein. The information at time n is usually larger than the observed values y1:n⁠. This could be modelled by introducing a filtration Gn such that σ(Y1:n)⊂Gn and defining Pn as Pn(⋅)=P(Yn+1∈⋅∣Gn)⁠. Such a generalisation should have a little cost, as most results on c.i.d. sequences work for an arbitrary filtration Gn⁠. Most probably I miss something, but I have some doubts on Section 2.4.1. It is obviously tempting to assign Pn as the empirical measure. But, it does not work. In fact, if Pn is the empirical measure for every n≥1⁠, one obtains the trivial sequence Yn=Y1 a.s. for each n. *(abstract via OpenAlex)*

JRSSB Bayesian Methods Clinical Trials 讨论 / 评论 · 发表于 2023-08-27

Marta Catalano, Augusto Fasano, and Giovanni Rebaudo’s contribution to the discussion of ‘Martingale posterior distributions’ by Fong, Holmes and Walker

Marta Catalano, Augusto Fasano, Giovanni Rebaudo

摘要

We congratulate the authors for an insightful foundational contribution to the statistical literature. This work builds on well-established methodologies, such as predictive inference, Doob’s theorem, and conditional independent sequences, providing a unifying framework and a novel understanding of Bayesian uncertainty. A central role is played by the predictive characterisation of the random parameter in terms of observable (yet unobserved) quantities, which are regarded as the root of all statistical uncertainty. This brings to a new approach to inference, termed martingale posterior, which sheds light on interesting connections between Bayesian and frequentist statistics. Indeed, both leverage on an empirical distribution: the former builds it through the predictive distributions, the latter through independent and identically distributed samples. Importantly, martingale posteriors go beyond some common homogeneity assumptions in the data, such as infinite exchangeability. Moreover, this approach may offer practical advantages both in terms of prior elicitation and computations, as illustrated by the authors in some relevant scenarios. We believe that the breadth of this contribution will inspire several future research questions. Here, we restrict our attention to two aspects that we found particularly interesting. First, the authors underline that inference and prediction under a martingale posterior might depend on the order of the data, even when a natural order does not exist. Such dependence will vanish as the sample size increases, but can still be relevant for finite samples. To overcome such an issue, the authors suggest using the average of the predictive distribution over M random permutations of the sample (e.g. M=10⁠). We believe that, in such a scenario, it could be useful to define and study predictive rules that go beyond infinite exchangeability yet preserve finite exchangeability for any fixed sample size. Second, from a theoretical and modelling perspective, it is often relevant to establish frameworks that weaken the homogeneity assumption of infinite exchangeability, while still preserving well-defined limits as the sample size increases and tractable learning updates. The authors rely on an additional principle: the martingale predictive coherence. We believe that it could be of interest to relax this principle to study other classes of converging predictive rules, e.g. those that preserve exchangeability but are not Kolmogorov consistent. To conclude, we believe the work by E. Fong, C. Holmes, and S. Walker will spur several new theoretical, modelling, and computational research directions. We commend the authors one more time for an outstanding paper. *(abstract via OpenAlex)*

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-08-27

Filippo Ascolani, Antonio Lijoi, and Igor Prünster’s contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Filippo Ascolani, Antonio Lijoi, Igor Prünster

摘要

We would like to congratulate the authors on their fine and insightful contribution, which provides an original perspective on Bayesian inference and opens up new exciting research directions. The key point is a novel interpretation of the role of prediction: in a Bayesian framework all the uncertainty lies in the unobserved data Yn+1,∞ and, once they are imputed through the predictive distributions, inference is straightforward. Therefore, prediction rules, besides allowing forecasting and extrapolation, are crucial also to infer parameters of interest. In Bayesian non-parametrics there, is a large stream of works focussing on the m-step ahead prediction for exchangeable species sampling data (see, e.g. Favaro et al., 2009; Lijoi et al., 2007), with recent contributions also in the partially exchangeable set-up (Camerlenghi et al., 2017). However, the predictive distributions are always determined through an indirect procedure that relies on the specification of a non-parametric prior and derives the prediction rule as a posterior expected value. The authors adopt a different, and more direct, approach by considering conditionally identically distributed sequences (Berti et al., 2023) that are only asymptotically exchangeable: in this case, the predictive distributions are available in closed form, but predictions may depend on the order of the observed data Y1:n⁠. This seems in contrast with the assumption of independent and identically distributed data that should imply invariance of inferential results with respect to permutation of the observations. Therefore, one may wonder whether the analysis could be extended so to come up with novel exchangeable predictives, without explicit reference to an underlying prior. To overcome the lack of invariance with respect to the ordering of the data, the authors suggest to average the predictions over different permutations of Y1:n⁠. While being computationally unfeasible, averaging over all the permutations induces a symmetry condition which is reminiscent of exchangeability. It would be interesting to check whether the ensuing prediction mechanism actually identifies an exchangeable sequence. This would boil down to showing invariance of the two-step ahead predictives (Fortini et al., 2000). If this were actually the case, the natural goal would be to identify the underlying prior. In general, the standard prior-likelihood mechanism may still be a plus when it comes to describing the dependence structure among the observations (i.e. the generative model). For instance, hierarchical models, that distinguish global and group-specific parameters, have proven to be useful in multiple fields. We wonder whether it would be possible to encode these structures directly into the predictive distributions: in particular, it would be interesting to ascertain whether the neat recursive expressions of Section 4, that allow fast computations, can be retained. *(abstract via OpenAlex)*

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-08-27

Bertrand Clarke's contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Bertrand Clarke

摘要

To see the effect of the paper on statistical practice, consider the following hypothetical dialogue between a Quantitative Scientist (QS) and a Predictive Statistician (PS). QS: I’ve been thinking about the data generating mechanism and what we can safely assume pre-experimentally. PS: Actually, I’m more interested in what you think your n+1 data point would be. If you were going to predict your n+1 measurement, how would you do it? QS: I’d guess what the true model is and take its mean or median. But, I don’t know what the true model is. PS: But you do know what the predictive density m(yn+1|y1:n) is, right? QS: If I assume my likelihood and prior are right. But, I don’t know that. PS: Do you believe there is an overall probability model for Y1:n+1? QS: Sure: M. PS: Then, we can use De Finetti to backform a likelihood and prior if we want. But let’s not. Let’s think about the martingale model m(yn+1|y1:n)⁠, m(yn+2|y1:n+1) and so on. QS: You mean what on average a new random draw would look like given the past data? If I’d seen y1:n+k I’d draw Yn+k+1 from some m(⋅|y¯n+k)⁠. I’d start with a density shaped like this [draws a density] and let it tighten a bit as n increases. PS: So that’s step k in your martingale model. We could use it, but let’s not. Instead, I’m going to form one-step ahead predictive DF’s using y1:n⁠, y1:n+1⁠, and so forth. I’m going to draw one data point from each DF to find the next. For each step up to N, I’ll find a version of the posterior mean. I’ll do this many times and then use the final posterior means to estimate W(⋅|y1:n)⁠. QS: You’re using repeated generation of data I don’t have to find the posterior I do have? PS: Yes, I’m completing your experiment conceptually by filling out your missing data. Over and over. Then, I can quantify the information you don’t have and derive the variability left in data you do have i.e. form the posterior. QS: And you’re ignoring everything we know about the data generator besides the data. Where does Pθ come in? PS: It doesn’t. We’re relying on martingale convergence in the mixture M. By changing the probability defining the mode of convergence we get the limits we want for Θ¯ and the DF—that’s why we can use them to estimate the posterior. QS: So, what role do the prior and likelihood—or even conditional mixtures—have anymore? If you’re right, why would we bother with them? Some sort of robustness analysis? PS: We could. But, we let’s not. Let’s think about the predictive process directly. QS: What’s the magic? PS: No magic. We just changed the mode from Pθ to M since we’re thinking about the whole countably infinite product space Y∞⁠. QS: Is θ still a parameter? PS: It’s a function value. Take Θn=Θn(Y1:n) and think of the posterior mean. Or better take n=∞⁠. Each Θ=θ has a Vθ={y1:∞|θ(y∞)=θ}⁠. QS: Don’t θ and a y1:∞ that I might conceptually get have to be compatible in that if y1:∞∉Vθ then θ can’t be true? PS: Yes. If you’re looking for magic, it’s that at infinity Θ=Θ(Y1:∞)=E(Θ|Y1:∞) pointwise in Y∞ and as random variables. Then the y1:∞ you get gives you your θ. So, in the limit, our computational procedure mimics what you would get if you had y1:∞ —but, we don’t. What you don’t sample gives an assessment of your uncertainty. QS: But I’m really using a sequence of predictive DF’s not posterior means or conditional predictives. PS: Yes. That way you get objectivity in your predictive thinking. *(abstract via OpenAlex)*

JRSSB Bayesian Methods 讨论 / 评论 · 发表于 2023-08-27

Ben Swallow's contribution to the Discussion of ‘Martingale Posterior Distributions’ by Fong, Holmes and Walker

Ben Swallow

摘要

Fong et al. (2023) present an interesting and novel approach to conducting Bayesian posterior estimation utilising a joint predictive distribution over missing potential observations and justify how this approach of generating one-step-ahead predictive distributions naturally aligns with the Bayesian likelihood-prior paradigm. The authors mention time series and hierarchical data as areas of potential future development. However, one of the major topics in Bayesian computational statistics is the problem of selecting a target model from a subset of candidate models or accounting for uncertainty across such models. In their paper, the authors present an approach based on a fixed model structures and conduct posterior inference within those model structures. Given the generality of the approach, I would be interested in hearing the authors’ comments on how this approach could be extended to the model uncertainty or model misspecification domain and to what extent the missing data dimension would need to scale with model space size in order to enable feasibility of any such approach. *(abstract via OpenAlex)*

SiM Latent Variables & Mixtures Bayesian Methods · 发表于 2023-08-27

A latent functional approach for modeling the effects of multidimensional exposures on disease risk

Sungduk Kim, Laura E. Beane Freeman, Paul S. Albert

摘要

Understanding the relationships between exposure and disease incidence is an important problem in environmental epidemiology. Typically, a large number of these exposures are measured, and it is found either that a few exposures transmit risk or that each exposure transmits a small amount of risk, but, taken together, these may pose a substantial disease risk. Further, these exposure effects can be nonlinear. We develop a latent functional approach, which assumes that the individual effect of each exposure can be characterized as one of a series of unobserved functions, where the number of latent functions is less than or equal to the number of exposures. We propose Bayesian methodology to fit models with a large number of exposures and show that existing Bayesian group LASSO approaches are a special case of the proposed model. An efficient Markov chain Monte Carlo sampling algorithm is developed for carrying out Bayesian inference. The deviance information criterion is used to choose an appropriate number of nonlinear latent functions. We demonstrate the good properties of the approach using simulation studies. Further, we show that complex exposure relationships can be represented with only a few latent functional curves. The proposed methodology is illustrated with an analysis of the effect of cumulative pesticide exposure on cancer risk in a large cohort of farmers.

JRSSB Latent Variables & Mixtures High-Dimensional Spatial & Time Series · 发表于 2023-08-24

Two-way dynamic factor models for high-dimensional matrix-valued time series

Chaofeng Yuan, Zhigen Gao, Xuming He, Wei Huang, Jianhua Guo

摘要

In this article, we introduce a two-way dynamic factor model (2w-DFM) for high-dimensional matrix-valued time series and study some of the basic theoretical properties in terms of identifiability and estimation accuracy. The proposed model aims to capture separable and low-dimensional effects of row and column attributes and their correlations across rows, columns, and time points. Complementary to other dynamic factor models for high-dimensional data, the 2w-DFM inherits the dimension-reduction feature of factor models but assumes additive row and column factors for easier interpretability. We provide conditions to ensure model identifiability and consider a quasi-likelihood based two-step method for parameter estimation. Under an asymptotic regime where the size of the data matrices as well as the length of the time series increase, we establish that the estimators achieve the optimal rate of convergence and are asymptotically normal. The asymptotic properties are reaffirmed empirically through simulation studies. An application to air quality data in Chinese cities is given to illustrate the merit of the 2w-DFM.

SMMR Nonparametric & Semiparametric · 发表于 2023-08-24

Smoothing Lexis diagrams using kernel functions: A contemporary approach

Philip S Rosenberg, Adalberto Miranda Filho, Julia Elrod, Aryana Arsham, Ana F Best, Pavel Chernyavskiy

摘要

Lexis diagrams are rectangular arrays of event rates indexed by age and period. Analysis of Lexis diagrams is a cornerstone of cancer surveillance research. Typically, population-based descriptive studies analyze multiple Lexis diagrams defined by sex, tumor characteristics, race/ethnicity, geographic region, etc. Inevitably the amount of information per Lexis diminishes with increasing stratification. Several methods have been proposed to smooth observed Lexis diagrams up front to clarify salient patterns and improve summary estimates of averages, gradients, and trends. In this article, we develop a novel bivariate kernel-based smoother that incorporates two key innovations. First, for any given kernel, we calculate its singular values decomposition, and select an optimal truncation point—the number of leading singular vectors to retain—based on the bias-corrected Akaike information criterion. Second, we model-average over a panel of candidate kernels with diverse shapes and bandwidths. The truncated model averaging approach is fast, automatic, has excellent performance, and provides a variance-covariance matrix that takes model selection into account. We present an in-depth case study (invasive estrogen receptor-negative breast cancer incidence among non-Hispanic white women in the United States) and simulate operating characteristics for 20 representative cancers. The truncated model averaging approach consistently outperforms any fixed kernel. Our results support the routine use of the truncated model averaging approach in descriptive studies of cancer.

JRSSB Latent Variables & Mixtures Bayesian Methods High-Dimensional · 发表于 2023-08-24

Karl Rohe and Muzhe Zeng’s reply to the Discussion of ‘Vintage factor analysis with varimax performs statistical inference’

Karl Rohe, Muzhe Zeng

摘要

Statistics as a field has been tremendously successful in creating and propagating the quantitative theories, techniques, and tools to do quantitative science. Because of our success, our community is continually fragmented into methodological subdisciplines within other fields1 , each a certain type of universe in which methodologies are continuously evolving (almost independently) in parallel. This phenomenon has happened slowly over time and as a result, we think it is time to reappraise the role of Statisticians with a capital S (i.e. in Statistics departments). In this vein, the history of Varimax provides a parable. Statisticians often perceive of our field as producing methodology by deriving it from our foundational theories (e.g. Maximum Likelihood, Bayesian, etc.). And then, other fields consume our methodologies. One issue with this perception is that the methodological subdisciplines are producing their own methodologies; what about the Statisticians with a capital S? Sometimes our foundational theories help them. Sometimes our theories do not help them. Sometimes their theories go beyond our own and then we call them Statisticians (with a capital S). Regardless, these are all skilled craftspeople and we fool ourselves when we pretend that Statisticians are the all-knowing producers. The history of factor analysis is an antidote to our illusions of grandeur. In 1935, Thurstone (a psychologist) was inspired by the idea of multiple types of intelligence and set out to create a way to measure them. Inspired by this and without an ‘electronic’ computer, he whittled a blunt tool, a way of iteratively plotting data and cleverly picking a tiny bit of a rotation and then iterating again. This was not derived from our theories. Moreover, this process was ‘subjective’ at every iteration. In 1956, Anderson and Rubin wrote a Gaussian theory that discounted factor rotations; their theory appears in every textbook on Multidimensional Statistics. But already in 1953, Carroll introduced an optimisation problem to ‘objectively’ pick a rotation using 4th-order moments. Then, Varimax came in 1958. Despite the protesting Statisticians, psychologists used factor rotations and conveyed them to future generations because factor rotations solved a problem. They did not need a ‘theoretical foundation’ and they should not need one now. We hope that the ‘theoretical foundation’ we have started to provide is that it might convince researchers in other disciplines to try using factor rotations on their problems. This parable of factor analysis is extreme because Statisticians have opposed it for nearly 90 years. Our fundamental claim is more general. Successful methodologies, the ones that spread, will have a consilience of a product-market-fit, ‘statistical theory’ (of models, theorems, and algorithms), and practical-know-how. In the successful diffusion of a methodology, Statistics as a discipline has an essential role. Ideally, we would be in a position to develop methodologies with product-market-fit, but often we are not. Our primary strength is that we can (1) provide a ‘statistical theory’, and then (2) leverage our central position in the academic network of methodological subdisciplines to convey this methodology and statistical theory. Before (1) developing a theory and (2) propagating it, there is a step (0). We believe Statisticians need to get better at step zero for our field to continue flourishing. The zero-problem is this: which methodologies should we support? Where should we direct our attention? Our journals are (unfortunately) filled with methodologies that lack a product-market-fit. There is a direct path to ensuring the methodology is fit: we can learn from others. This alternative path leverages and reinforces our central role in the network of quantitative subdisciplines. When we learn popular techniques from others, they already have product-market-fit and likely already have practical-know-how. Our job is to give it a model and a theory (broadly interpreted) and to make it into something that other researchers might enjoy. Maybe this also inspires new algorithms and new estimators. Maybe it does not. By developing this framework, we enable other fields to learn. Importantly, this is not an entirely new way of doing Statistics; its the way it is already happening. In this process, methodologies are not derived, but rather methodologies evolve. We Statisticians can play a fundamental role in methodologies evolving and reaching consilience, but we should stop assuming that product-market-fit is easy. Instead, we should recognise the more realistic role that we play in the evolution of quantitative methodologies and leverage the subdisciplines that are simultaneously and endlessly refining numerous different methodologies and testing their product-market-fit. Varimax is a parable for this point. Thank you to the Discussion Meetings Committee for hosting this discussion. Thank you Dr. C. Grazian for chairing the meeting. Thank you Professors Hoff and Pensky for leading the discussion. Thank you to the discussants for your thoughtful and inspiring comments. It was an honour to take part in this venerable tradition. We are excited for the interest in Varimax. Many of the discussants highlight additional areas that need more exploration. We agree. There is so much more work to be done. We will use the rest of this rejoinder to find common themes for future directions and emphasise some ways that we might learn from the subdisciplines to pursue these directions. We aim bring together some key threads in this discussion; as such, many excellent and interesting questions and comments from the discussants are not captured here. In places where we express scepticism, we do not say this to discourage any research in these directions, but instead to identify the key challenges that we suspect will need to be addressed with various approaches. The basis for Principal Components Analysis (PCA) comes from the singular value decomposition (SVD), which provides the optimal low-rank approximation for a data matrix in a least-squares sense. We show that Varimax can pick new axes for this subspace that are reasonable under a broad class of models. However, Varimax is certainly not the only way to choose the axes (i.e. rotate the PCs). Kaizheng Wang gives a counter example where Varimax fails because the natural (sparse) factor coordinates are platykurtic. Similarly, Bickel, Bean, Chen, and Sarkar highlight that our theorem requires Leptokurtosis, a specific kind of non-Gaussianity. Some Independent Components Analysis (ICA) objectives can recover natural coordinates in other or perhaps more general conditions. In our limited experience with social networks, there is always degree heterogeneity and this heterogeneity always manifests with radial streaks in the pairs plot (e.g. Figure 2 in the discussed paper). These radial streaks are a key diagnostic for leptokurtic factors. In such situations, in our experience, other rotation techniques tend to create local optima that are not empirically beneficial. Bickel, Bean, Chen, and Sarkar mention an assumption from the literature on Non-Negative Matrix Factorisation about ‘pure nodes or anchor words’ (Donoho & Stodden, 2003). Under this identifying assumption, for each column in Z, there must be a row of Z that only loads in this column. While the existence of such pure nodes will likely (approximately) exist under independent leptokurtic factors in fixed dimension k, finding them might be tricky. Moreover, aesthetically, it seems not as satisfying as an objective function (like Varimax) that smoothly incorporates all data. In fact, it was precisely the assumption of ‘pure nodes’ that led us to search for smoother alternatives; we then happened upon Varimax. It is tempting to consider new types of rotations and we do not wish to discourage that. In addition, we hope that we can also direct our attention to unifying the disjoint literatures on rotations. The literature on ICA has developed theory and algorithms that are comparatively underdeveloped in the literature on factor analysis. At the same time, the deeper theoretical understandings in the literature on Independent Component Analysis rests primarily on the heuristic of non-gaussianity, which seems comparatively underdeveloped to the heuristic notions of sparsity, radial streaks, and simple structure in the literature on factor analysis. To re-emphasise a point of the paper, Varimax has survived for close to 70 years despite strong pressure against it. We suspect that this survival is linked to its empirical fitness. What is it about Varimax that make researchers tend to prefer it to other rotations? Perhaps ‘leptokurtosis’ is a better assumption? Alternatively, perhaps Varimax is better behaved algorithmically. Perhaps, it has better statistical performance in the presence of noise? Perhaps it is the ‘Kaiser normalisation’ that we do not account for in our paper? It is possible that we will discover more beneficial rotations for certain settings and as theoreticians, we are all drawn to that prospect. It is also important to further understand the successes that have existed over the past 90 years and ensure our theory captures those success as fully as possible. This point aligns with the first paragraphs of this rejoinder. Going forward, a key aspect of this theory could be a more unifying lens to understand rotations and their relationships. In this vein, Siroki and Strimmer ask a brilliant question that unifies the spirit of ICA with the algorithms of factor analysis: ‘[is the] leptokurtic condition strict enough and sufficient to single out Varimax or whether perhaps the whole orthomax family may be compatible with it?’ We do not know, but we are curious and hope that someone will find out! For a path forward, it will likely be helpful that Chu and Trendafilov (1998) give the first- and second-order conditions for Orthomax. We hope for more inquiries that bring together the richness of understanding that we have from 90 years of factor rotations and 30 years of ICA. Vintage factor analysis is a powerful tool for uncovering patterns of association, but we must still go through the difficult interpretative work of trying to assess why the patterns of correlation are present, and what they mean. VanderWeele argues that the meaning of factors extends beyond anything inside the factor analysis; it is a necessary post-hoc step to give them meaning. We agree. He continues that Thurstone’s notion of a ‘Scientifically Meaningful Category’ is ambiguous. We strongly agree. Speaking to our analysis of scientific publishing, he writes ‘There are various ways specific disciplines might be precisely defined, with varying levels of scope, but this then is again a conceptual, rather than an empirical, question’. Here, we partially disagree. Whenever we seek to cut nature at the seams (e.g. perform clustering or factor analysis), it is important to know the purpose of cutting; identifying our purpose is clearly conceptual. That said, once we understand our purpose, we see no reason to exclude empirical evidence from such tailoring. Yang Liu discusses the classical framing of ‘sense making’ with rotations. In particular, multiple different rotations can all make sense and can have different types of sparsity. Liu presents an interesting counter example from The development of hierarchical factor solutions (Schmid & Leiman, 1957), where the second solution is rank deficient, but has a convenient hierarchical interpretation. This suggests a connection to hierarchical clustering. Liu was not the only discussant to highlight a connection to hierarchies. Cai, Yang, Zhao, and Zhu propose using Varimax within previous approaches to hierarchical clustering. Moreover, VanderWeele correctly critiques our analysis and the semi-parametric factor model for lacking hierarchical structure; ‘the non-hierarchical nature of the categories as the number of factors increases (Section 3.1.3), indicates that while these are potentially interpretable categories, they are not fixed scientific realities’. Here, there is room for the theory and methods of hierarchies and factor analysis to hybridise and grow. Unbeknownst to most Statisticians, a great deal of statistical theory for hierarchical clustering has been developed in the literature on phylogenetic tree reconstruction. In that literature, ‘hierarchical clustering’ is not simply an exploratory technique. Rather, it deeply informed by the theories and models of evolution. This backdrop has provided an environment with strong selection pressures for the fittest hierarchical clustering methodologies. We should all seek to incorporate what they have learned, identify the parts that likely generalise to other fields, and hybridise it with techniques from other subdisciplines. In a forthcoming manuscript, Sijia Fang and co-author Rohe study one such relationship in hierarchical modelling of social networks. They propose the T-Stochastic Graph model; vsp can be used to identify part of the hierarchy T⁠. We hope to see much more inquiry into the relationship between factoring and hierarchies and much more inquiry into hierarchical clustering. Marianna Pensky’s simulations suggest that in certain settings, more fully (and still correctly) specified model fitting techniques can out-perform vsp. We imagine vsp as one step in a fitting pipeline. Even in situations where one does not use vsp in the final estimates, it can (1) provide the opportunity to first diagnose what structure the latent space appears to have and then, potentially and (2) provide a statistically consistent initialisation for a more refined fitting procedure. In particular, we hope that this insight could be built into the choice of latent space prior for Bayesian approaches as Xiaoyue Niu discusses. This choice of prior can be diagnosed by inspecting the pairs plots (i.e. Figure 2 in the discussed paper). What other models can we better understand with rotations? We are optimistic that Statisticians will become more involved in making sense of the black box Large Language Models (LLMs) such as the GPTs; within these models, there are ‘low-dimensional’ embeddings with K>700⁠. These essentially have K>700 factors, where the individual dimensions have not yet been interpreted. These dimensions were not computed with SVD, but rather with Stochastic Gradient Descent and a far more complicated objective functions. But the same questions remain. What do these dimensions mean? Do they display radial streaks (we hypothesise that they do)? If so, perhaps one component of that sense making might be factor rotations and hierarchies. Alexander Van Werde asks ‘Are there situations where one should be careful [applying the theory of the current paper]?’ In our experience, the biggest obstacle to successfully using PCA and Varimax in empirical applications is the problem of localisation, something that is drastically under discussed in any related literature. Localisation can be diagnosed. Look at the pairs plots of your Varimax factors (i.e. Figure 2 in the discussed paper). Do you see any axes that do not have a streak, but instead look like a one to three outliers? If so, then we say that the component has localised. It might have identified a very important data point/measurement. Or, it could be an artefact of noise. Either way, it is irritating to dedicate an entire dimension to identifying a few measurements. Sometimes, all of your dimensions will look this way and you might need to dramatically increase k to find meaningful factors. In classical factor analysis, there are ‘Heywood cases’ and in neural networks there is ‘neural collapse’; in an heuristic sense, we think of these three things as similar types of failures of three distinct techniques. Chen and Xu combine vsp into one objective function. However, there is an extreme tension between PCA and Varimax that is not discussed in the paper. PCA can localise, with extreme outliers in one element (or perhaps a few elements). Such localised dimensions have a massive kurtosis. So, both PCA and Varimax will enjoy finding these dimensions. By combining PCA and Varimax into a single objective, we worry that it would amplify these failures. Le et al. (2017) provide the first improved bounds for spectral convergence with regularisation. Ke and Wang (2022) use the phrase ‘pre-PCA normalisation’ and discuss the benefits of normalisation and regularisation. Zhang and Rohe (2018) illustrate the types of patterns in random graphs that generate localisation and how regularisation addresses this issue. More work is needed in these directions. In particular, there is relatively little methodological and applied work that could communicate most directly with newcomers to PCA with high dimensional data. Han and Zhang provide a large step forward in our understanding of the Varimax solution to U when the dimension K increases. There is room to explore how it behaves with U^⁠. In a related vein, we suspect that the bound in our main theorem can and will be improved. This is not something that we are currently pursuing. Joshua Cape hints at a result on the asymptotic normality of the rows of Z^⁠. We hope that others will join them in exploring these direction. *(abstract via OpenAlex)*

SiM Networks & Graphs Testing & Uncertainty · 发表于 2023-08-24

A two‐step log‐linear procedure for graphical representation and inference of associations in cross‐classified data for disease diagnosis

J. Fernando Vera, José A. Roldán‐Nofuentes

摘要

Biometrical sciences and disease diagnosis in particular, are often concerned with the analysis of associations for cross‐classified data, for which distance association models give us a graphical interpretation for non‐sparse matrices with a low number of categories. In this framework, usually binary exploratory and response variables are present, with analysis based on individual profiles being of great interest. For saturated models, we show the usual linear relationship for log‐linear models is preserved in full dimension for the distance association parameterization. This enables a two‐step procedure to facilitate the analysis and the interpretation of associations in terms of unfolding after the overall and main effects are removed. The proposed procedure can deal with cross‐classified data for profiles by binary variables, and it is easy to implement using traditional statistical software. For disease diagnosis, the problems of a degenerate solution in the unfolding representation, and that of determining significant differences between the profile locations are addressed. A hypothesis test of independence based on odds ratio is considered. Furthermore, a procedure is proposed to determine the causes of the significance of the test, avoiding the problem of error propagation. The equivalence between a test for equality of odds ratio pairs and the test for equality of location for two profiles in the unfolding representation in the disease diagnosis is shown. The results have been applied to a real example on the diagnosis of coronary disease, relating the odds ratios with performance parameters of the diagnostic test.

Biometrics Survival Analysis Clinical Trials · 发表于 2023-08-23

Study Design for Restricted Mean Time Analysis of Recurrent Events and Death

Lu Mao

摘要

The restricted mean time in favor (RMT-IF) of treatment has just been added to the analytic toolbox for composite endpoints of recurrent events and death. To help practitioners design new trials based on this method, we develop tools to calculate the sample size and power. Specifically, we formulate the outcomes as a multistate Markov process with a sequence of transient states for recurrent events and an absorbing state for death. The transition intensities, in this case the instantaneous risks of another nonfatal event or death, are assumed to be time-homogeneous but nonetheless allowed to depend on the number of past events. Using the properties of Coxian distributions, we derive the RMT-IF effect size under the alternative hypothesis as a function of the treatment-to-control intensity ratios along with the baseline intensities, the latter of which can be easily estimated from historical data. We also reduce the variance of the nonparametric RMT-IF estimator to calculable terms under a standard set-up for censoring. Simulation studies show that the resulting formulas provide accurate approximation to the sample size and power in realistic settings. For illustration, a past cardiovascular trial with recurrent-hospitalization and mortality outcomes is analyzed to generate the parameters needed to design a future trial. The procedures are incorporated into the rmt package along with the original methodology on the Comprehensive R Archive Network (CRAN).

SiM Clinical Trials Bayesian Methods · 发表于 2023-08-23

A comparison of Bayesian information borrowing methods in basket trials and a novel proposal of modified exchangeability‐nonexchangeability method

Libby Daniells, Pavel Mozgunov, Alun Bedding, Thomas Jaki

摘要

Recent innovation in trial design to improve study efficiency has led to the development of basket trials in which a single therapeutic treatment is tested on several patient populations, each of which forms a basket. In a common setting, patients across all baskets share a genetic marker and as such, an assumption can be made that all patients may have a homogeneous response to treatments. Bayesian information borrowing procedures utilize this assumption to draw on information regarding the response in one basket when estimating the response rate in others. This can improve power and precision of estimates particularly in the presence of small sample sizes, however, can come at a cost of biased estimates and an inflation of error rates, bringing into question validity of trial conclusions. We review and compare the performance of several Bayesian borrowing methods, namely: the Bayesian hierarchical model (BHM), calibrated Bayesian hierarchical model (CBHM), exchangeability‐nonexchangeability (EXNEX) model and a Bayesian model averaging procedure. A generalization of the CBHM is made to account for unequal sample sizes across baskets. We also propose a modification of the EXNEX model that allows for better control of a type I error. The proposed method uses a data‐driven approach to account for the homogeneity of the response data, measured through Hellinger distances. Through an extensive simulation study motivated by a real basket trial, for both equal and unequal sample sizes across baskets, we show that in the presence of a basket with a heterogeneous response, unlike the other methods discussed, this model can control type I error rates to a nominal level whilst yielding improved power.

SiM Survival Analysis · 发表于 2023-08-22

Estimation of the time of exposure based on interval and censored data using theε‐accelerated EM algorithm

Daisuke Yoneoka, Takayuki Kawashima, Yuta Tanoue, Shuhei Nomura, Akifumi Eguchi

摘要

Accurately estimating the timing of pathogen exposure plays a crucial role in outbreak control for emerging infectious diseases, including the source identification, contact tracing, and vaccine research and development. However, since surveillance activities often collect data retrospectively after symptoms have appeared, obtaining accurate data on the timing of disease onset is difficult in practice and can involve “coarse” observations, such as interval or censored data. To address this challenge, we propose a novel likelihood function, tailored to coarsely observed data in rapid outbreak surveillance, along with an optimization method based on an ‐accelerated EM algorithm for faster convergence to find maximum likelihood estimates (MLEs). The covariance matrix of MLEs is also discussed using a nonparametric bootstrap approach. In terms of bias and mean‐squared error, the performance of our proposed method is evaluated through extensive numerical experiments, as well as its application to a series of epidemiological surveillance focused on cases of mass food poisoning. The experiments show that our method exhibits less bias than conventional methods, providing greater efficiency across all scenarios.

SiM Genetics & Omics · 发表于 2023-08-22

A flexible quasi‐likelihood model for microbiome abundance count data

Yiming Shi, Huilin Li, Chan Wang, Jun Chen, Hongmei Jiang, Ya‐Chen T. Shih, Haixiang Zhang, Yizhe Song, Yang Feng, Lei Liu

摘要

In this article, we present a flexible model for microbiome count data. We consider a quasi‐likelihood framework, in which we do not make any assumptions on the distribution of the microbiome count except that its variance is an unknown but smooth function of the mean. By comparing our model to the negative binomial generalized linear model (GLM) and Poisson GLM in simulation studies, we show that our flexible quasi‐likelihood method yields valid inferential results. Using a real microbiome study, we demonstrate the utility of our method by examining the relationship between adenomas and microbiota. We also provide an R package “fql” for the application of our method.

SiM Diagnostics & Prediction · 发表于 2023-08-21

Approximate likelihood and pseudo‐likelihood inference in meta‐analysis of diagnostic accuracy studies accounting for disease prevalence and study design

Annamaria Guolo

摘要

Bivariate random‐effects models represent a recommended approach for meta‐analysis of diagnostic test accuracy, jointly modeling study‐specific sensitivity and specificity. As the severity of the disease status can vary across studies, a proper analysis should account for the dependence of the accuracy measures on the disease prevalence. To this aim, trivariate generalized linear mixed‐effects models have been proposed in the literature, although computational difficulties strongly limit their applicability. In addition, the attention has been mainly paid to cohort studies, where the study‐specific disease prevalence can be estimated from, while information from case‐control studies is often neglected. To overcome such limits, this article introduces a trivariate approximate normal model, which accounts for disease prevalence along with accuracy measures in cohort studies and sensitivity and specificity in case‐control studies. The model represents an extension of the bivariate normal mixed‐effects model originally developed for meta‐analysis not accounting for disease prevalence, under an approximate normal within‐study distribution for the logit of estimated sensitivity and specificity. The components of the approximate within‐study covariance matrix are derived and the likelihood function is obtained in closed‐form. The approximate likelihood approach is compared to that based on the exact within‐study distribution and to its modifications following a pseudo‐likelihood strategy aimed at reducing the computational effort. The comparison is based on simulation studies in a variety of scenarios, and illustrated in a meta‐analysis about the accuracy of a test to diagnose fungal infection and a meta‐analysis of a noninvasive test to detect colorectal cancer.

JASA Causal Inference Econometrics & Economics · 发表于 2023-08-21

An Automated Approach to Causal Inference in Discrete Settings

Guilherme Duarte, Noam Finkelstein, Dean Knox, Jonathan Mummolo, Ilya Shpitser

摘要

Applied research conditions often make it impossible to point-identify causal estimands without untenable assumptions. Partial identification-bounds on the range of possible solutions-is a principled alternative, but the difficulty of deriving bounds in idiosyncratic settings has restricted its application. We present a general, automated numerical approach to causal inference in discrete settings. We show causal questions with discrete data reduce to polynomial programming problems, then present an algorithm to automatically bound causal effects using efficient dual relaxation and spatial branch-and-bound techniques. The user declares an estimand, states assumptions, and provides data-however incomplete or mismeasured. The algorithm then searches over admissible data-generating processes and outputs the most precise possible range consistent with available information-that is, sharp bounds-including a point-identified solution if one exists. Because this search can be computationally intensive, our procedure reports and continually refines non-sharp ranges guaranteed to contain the truth at all times, even when the algorithm is not run to completion. Moreover, it offers an ε -sharpness guarantee, characterizing the worst-case looseness of the incomplete bounds. These techniques are implemented in our Python package, autobounds. Analytically validated simulations show the method accommodates classic obstacles-including confounding, selection, measurement error, noncompliance, and nonresponse. Supplementary materials for this article are available online. *(abstract via PubMed)*

SiM Spatial & Time Series High-Dimensional Nonparametric & Semiparametric · 发表于 2023-08-20

Shrinkage estimators of the spatial relative risk function

Martin L. Hazelton

摘要

The spatial relative risk function describes differences in the geographical distribution of two types of points, such as locations of cases and controls in an epidemiological study. It is defined as the ratio of the two underlying densities. Estimation of spatial relative risk is typically done using kernel estimates of these densities, but this procedure is often challenging in practice because of the high degree of spatial inhomogeneity in the distributions. This makes it difficult to obtain estimates of the relative risk that are stable in areas of sparse data while retaining necessary detail elsewhere, and consequently difficult to distinguish true risk hotspots from stochastic bumps in the risk function. We study shrinkage estimators of the spatial relative risk function to address these problems. In particular, we propose a new lasso‐type estimator that shrinks a standard kernel estimator of the log‐relative risk function towards zero. The shrinkage tuning parameter can be adjusted to help quantify the degree of evidence for the existence of risk hotspots, or selected to optimize a cross‐validation criterion. The performance of the lasso estimator is encouraging both on a simulation study and on real‐world examples.

第 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 页